AWS Launches Web Answer for Cited Web-Grounded Responses
Summary:AWS has announced the general availability of Web Answer on Amazon Bedrock AgentCore. This managed connector accepts a natural-language question, plans and executes parallel searches, reads key pages to verify facts, and searches again when the evidence is insufficient. It then returns a synthesized answer whose claims are linked to source URLs through a single Model Context Protocol tool call. The service packages search, page retrieval, synthesis, and citation attribution into an AWS-operated
A Cited Web-Grounded Answer in One Tool Call
AWS has made Web Answer generally available on Amazon Bedrock AgentCore. A developer submits a natural-language question, and the managed service plans its own searches, executes them in parallel, reads key pages to verify facts, and searches again if the evidence is insufficient. It returns a synthesized answer with source URLs as a Model Context Protocol tool result.
Web Answer is not a new foundation model. It is a managed retrieval connector that packages search, page retrieval, fact checking, answer synthesis, and citation attribution into an AWS-operated service.
How the Retrieval Loop Works
A typical request follows this path:
- The agent invokes the tool through an AgentCore Gateway over MCP.
- The Gateway authenticates the caller with AWS IAM or a JWT.
- The Gateway calls the AWS-owned Web Answer connector with its service role.
- Web Answer creates a search plan and runs multiple searches in parallel.
- It reads key pages and checks important facts, searching again when the evidence is weak.
- It synthesizes the answer and maps individual claims to source URLs.
- The Gateway returns the answer, citations, and metadata to the agent.
The important distinction is the model-driven iterative retrieval loop. Search planning, query expansion, page selection, fact verification, and stopping decisions happen inside the managed service rather than in application code.
Web Answer performs new searches for every question and supplies the model with the current date to reduce stale-index errors. It also supports publication-date ranges, country localization, and domain filters. Administrators define target-level domain boundaries, while an individual call may narrow those boundaries but cannot widen them.
How It Differs from Web Search
Both tools support retrieval, but they divide responsibility differently.
| Capability | Web Search | Web Answer | |---|---|---| | Output | Ranked URLs, titles, snippets, and metadata | A synthesized answer with citations | | Retrieval | Caller processes the results | Service runs an agentic search loop | | Page reading | Caller implements it | Service reads key pages and verifies facts | | Synthesis | Caller's model performs it | Web Answer performs it internally | | Attribution | Caller must engineer it | Service maps claims to source URLs | | Best fit | Broad research and custom result processing | Fast factual Q&A within a larger agent |
They can be combined. An agent may use Web Search for broad research that requires raw evidence, then call Web Answer for a specific factual subtask that benefits from a ready-made cited response.
Integration and Permission Boundaries
AWS owns and operates Web Answer. A Gateway administrator must grant the Gateway service role bedrock-agentcore:InvokeWebAnswer on the AWS-managed web-answer.v1 ARN. The developer then creates a Gateway target with the web-answer connector ID and selects GATEWAY_IAM_ROLE for outbound credentials.
This design separates inbound and outbound identities. The Gateway's inbound authorizer determines who can invoke it, while the service role determines how the Gateway reaches Web Answer. The Gateway role does not need bedrock:InvokeModel; the agent's model continues to run under the agent's own identity.
Once attached, agents discover and invoke the tool through the standard MCP tools/list and tools/call methods. The same mechanism works with Strands Agents, LangChain, LangGraph, CrewAI, and other MCP-compatible clients.
Structured Responses and Operational Value
The response contains both textual content and structuredContent. The answer is Markdown with inline citation markers. The structured payload separately provides the answer text and a citation list containing URLs and, when available, domain names.
This dual representation is useful in production. A model can consume the inline citations while backend services use the structured data for source auditing, domain statistics, and caching. The response metadata also includes latencyMs, although the official example represents a single observation rather than a latency distribution or service guarantee.
Citation coverage alone does not prove factual correctness. A citation shows where the system claims an answer came from; it does not prove that the source is authoritative, that the quoted passage supports the claim, or that the model interpreted the evidence correctly. Production evaluations should still measure claim-level citation accuracy, source quality, claim-evidence consistency, and behavior when sources conflict.
Latency, Cost, and Control Trade-offs
Because Web Answer combines multiple searches, page reads, fact checks, and generation into one managed loop, its end-to-end latency is normally higher than a Web Search call that only returns ranked links. The official sample reports 5,210 milliseconds, but that is neither a guarantee nor representative of every query class.
Its main benefit is lower operational complexity, not faster individual requests. Teams no longer need to maintain their own crawlers, page cleaners, rerankers, evidence selectors, prompts, citation parsers, and retry logic. The tradeoff is reduced visibility into some retrieval internals and less control over search backends, internal models, prompting, and intermediate evidence selection.
The listed price is $10 per 1,000 tool calls, or $0.01 per call. Gateway invocation charges and the cost of the agent's model should be checked separately in AWS billing. For frequently asked, fixed questions, application-level caching can avoid repeating the entire retrieval cycle.
AWS also supports VPC interface endpoints and AWS PrivateLink, allowing AgentCore Gateway traffic to remain within an enterprise network boundary. CloudWatch receives operational metrics from the first invocation. The per-account limit is 10 transactions per second in us-east-1, with increases available through AWS Support.
Evaluation Requirements
The announcement does not disclose standardized figures for answer accuracy, citation correctness, abstention rate, or comparison with alternative retrieval systems. Production acceptance tests should cover:
- Claim-level citation accuracy rather than the mere presence of URLs;
- Correct abstention when information is outdated, conflicting, or unsupported;
- Authoritative-source coverage and the rate of links outside approved domains;
- Resistance to malicious pages and prompt injection embedded in retrieved content;
- Fallback behavior for search failures, inaccessible pages, broken citations, and timeouts;
- P50, P95, and P99 latency by query difficulty, together with retry cost.
For financial, medical, and legal workflows, human review or deterministic policy layers remain appropriate. A readable answer with citations is a retrieval aid, not automatically a final decision that satisfies regulatory requirements.
Availability and Adoption Guidance
Web Answer is currently available only in AWS US East (N. Virginia), or us-east-1. The walkthrough requires an IAM-authenticated AgentCore Gateway and Python 3.10 or later.
Use Web Search when the application needs raw results, a custom model-driven research process, or tighter control over the retrieval trace. Use Web Answer for bounded factual subtasks where a managed cited answer is more valuable than operating a separate search and attribution pipeline. In either case, caching, domain restrictions, citation validation, and explicit failure fallbacks remain important for production deployments.
— END —Source: AWS Machine Learning Blog (2026-11-05)
Tags: #Amazon Bedrock AgentCore #Web Answer #Agentic Search #MCP #RAG
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments